The present invention is directed to methods and software tools for producing Application-Specific Integrated Circuits (ASICs).
ASICs and Programmable Logic Devices (PLDS) represent two distinct types of Integrated Circuits (ICs) that perform specific logic operations. ASICs are distinguished from PLDs in that ASICs are customized during the design and fabrication process to implement a user""s logic function, while PLDs are provided with programmable circuitry that is configured (programmed) by a user after the PLD is fabricated.
FIG. 1 is a flow diagram illustrating a simplified conventional process for designing and fabricating an ASIC 100. At the beginning of the design phase, a circuit designer typically selects an ASIC family that is defined, for example, by the feature and chip size of the completed ASIC, along with any pre-defined arrangements of transistors and/or contact pads. The user then enters a logic design into a computer or workstation 110 using an input device 111. Computer 110 is loaded with a logic design entry software tool 113 that is typically linked to a design library 115 containing design parameters associated with the selected ASIC family. Design library 115 also typically includes one or more pre-defined logic elements (often referred to as xe2x80x9cmacrosxe2x80x9d) that can be selected by the user during the logic design entry process. The user utilizes entry tool 113 to link these pre-defined logic elements with custom-designed logic elements (if any) to construct the logic design. Once the logic design is entered, place and route tools 117, also loaded on computer 110, are utilized to generate a place-and-route solution, which arranges the various interrelated portions of the logic design in an efficient two-dimensional spatial relationship that can be fabricated on a chip, and generates signal (interconnect) lines for passing signals between the interrelated portions. A layout tool 119 is then utilized to construct a three-dimensional representation of the actual circuit structures (e.g., regions of doped semiconductor, insulated regions, and metal lines) needed to implement the logic design. Next, this three-dimensional representation is then used to generate a set of masks (step 120) that are then used to fabricate ASIC 100 (step 130) using known techniques.
Two popular techniques for building ASICs are Mask-Programmed Gate Arrays (MPGAs), Standard Cell or Cell Based Integrated Circuits (CBICs).
FIG. 2 is a perspective view showing a simplified MPGA 200, along with a mask set 210 used during the fabrication phase used to form MPGA 200. MPGA 200 includes rows of pre-defined transistors 202 and surrounding bonding pads 205 that are laid-out independent of the design to be manufactured. Specifically, mask set 210 includes a first group of pre-defined masks 212 that are utilized during the fabrication of all ASICs xe2x80x9cbuiltxe2x80x9d on MPGA 200. Note that during the ASIC design phase (described above), the placement process assigns portions of the logic function to selected pre-defined transistors 202. An ASIC is xe2x80x9cbuiltxe2x80x9d on pre-defined transistors 202 using custom mask set 214 that produces interconnect wiring 206 between the assigned transistors 202 to perform the logic functions of the ASIC logic design. Because the layout of transistors 202 is the same for all designs, mask set 212 (i.e., the masks used to fabricate transistors 202) can be shared among all ASIC designs of a particular MPGA family, thereby reducing the costs and shortening the time needed to implement ASIC logic designs by reducing the number of custom masks. Further, the production of custom masks is facilitated using wiring patterns stored in design library 115 (see FIG. 1) that are provided for producing often-used circuit structures, such as logic gates. Typically, design library 115 provides lower level metal layer patterns that wire together predefined sets of transistors to build logic gates. The logic gates are interconnected by longer wires 206 using upper level metal layers to build the ASIC logic design. These upper level metal layers are defined by place and route tools 117. The use of such a design library further reduces the costs and shortens the time needed to implement ASIC designs by automating much of the custom mask design process.
In contrast to MPGAS, CBICs are essentially blank silicon substrates upon which both transistors and wiring connections are custom-made for each design. Similar to MPGAs, certain commonly-used structures (e.g., the transistors and wiring connections of logic gates) are selected from a design xe2x80x9clibraryxe2x80x9d, and are selectively arranged in custom cells according to a desired ASIC logic design. Because even the transistors are customized in CBICS, designs can be much more complex, and include dense memory and custom data path structures, at the cost of more mask-making and the need to develop custom cells for the logic gates.
Modern IC processes have reduced the size of transistors such that hundreds of millions of transistors to be inexpensively fabricated on a single MPGA or CBIC chip. However, as the numbers of transistors on each chip increase, noise in the interconnect wiring connecting these transistors is becoming an increasing problem. In particular, as IC manufacturing processes improve, signal delay, particularly for xe2x80x9clong distancexe2x80x9d signals (i.e., signals transmitted between distinct groups of transistors), is dominated by the signal delay caused by transmission through the interconnect wiring. Further, as interconnect wiring gets narrower and thinner with advanced processes, its resistance increases, thereby further delaying the transmission of signals over longer interconnect wires.
One method to address signal delays in long interconnect wires is to segment each interconnect wire into collinear wire segments that are connected in series by buffers. That is, an interconnect wire that spans a long distance in an ASIC design may produce a greater signal delay than a buffered series of wire segments that are optimized to reduce signal delays. The optimal wire segment length (i.e., spacing between buffers) of these buffered signal paths depends primarily on the process technology (i.e., the resistance and capacitance the wire). As process technology advances, the correct fixed length gets smaller and smaller. Further, the sizing of the wire segment and the positions of the buffers depend on nearby wire segments. Therefore, the optimal wire segment and associated buffer size will be different for each family of ASICs that are defined by these features.
A problem with incorporating buffered interconnect wires into a conventional ASIC design processes is the difficult task of automatically breaking the longer wires into optimal-length wire segments, and providing necessary silicon space for the buffers. According to conventional ASIC design processes, the tasks of designing the segmented interconnect wires and placing buffer structures are performed by conventional placement and routing tools, mentioned above, which are highly-sophisticated software programs. Modification of these software programs to break all wires into optimal-length wire segments and to provide buffer space is impractical because the calculations are too difficult to perform quickly during placement and routing. Further, routing must be done after placement, but buffer insertion requires that they be placed, so if buffers are inserted during routing, the placement must change, potentially requiring the route to be re-done also. As a result, it is difficult to predict the actual delay of signals transmitted between circuit groups in an ASIC design prior to routing. This prediction is needed in placement and synthesis to meet timing requirements for the ASIC design.
What is needed is a method and structure that facilitates predictable and optimal-speed interconnect wiring for long distance signal transmissions in ASIC designs. Preferably this method and structure would permit sharing of masks to reduce costs, yet allow custom transistors for dense memory blocks.
The present invention is directed to a method for making ASICs that utilizes pre-engineered routing structures for implementing long-distance signal connections. The pre-engineered routing structures include sets of optimal-length collinear interconnect wire segments and optimal-size interlinking buffers that minimize the signal delay associated with long distance signal transmissions in each family of ASIC chips. Note that short-distance connections of the ASIC design, which do not require buffering, are implemented using conventional interconnect wiring. The pre-engineered routing structures are implemented during the place-and-route phase of ASIC design by providing necessary connections to source and destination logic elements of the ASIC. Because the interconnect wire segments are pre-defined, the masks used to make the wiring segments can be pre-made, sharing the mask-making costs for all ASIC designs utilizing a particular ASIC family.
In accordance with an aspect of the present invention, the optimal length of each interconnect wire segment is calculated for each ASIC family by simulating wire segments of various lengths using distributed resistance and capacitance wire models, and estimating crosstalk from neighboring wires. Similarly, the size of each buffer is designed to optimize the speed of signal transmission on each optimal-length wire segment. Because the routing structures are pre-engineered in advance for each family of ASIC chips, rather than once for every logic design to be implemented by the ASIC (or once for each trial route of a signal), signal delays can be calculated during the ASIC design process, thereby greatly simplifying the ASIC design process.
In accordance with another aspect of the present invention, a design library is provided for a family of ASICS that includes the pre-engineered routing structures. One or more upper metal layers of the ASIC production process are reserved in the library for implementing the optimal-length wire segment, and lower metal layers are utilized to facilitate short-distance signal transmissions and connections between adjacent pairs of collinear wire segments (i.e., by connecting to associated interlinking buffers). In one embodiment, a first metal layer is utilized to produce a first set of wiring segments extending in a first direction, and a second metal layer is utilized to produce a second set of wiring segments extending perpendicular to the first set. Space is reserved on the underlying substrate for the buffers of the pre-engineered routing structure. For efficiency, the buffers are arranged so that they can be connected to drive signals in either direction (e.g., from a first segment to a second segment, or from the second segment to the first segment). Adjacent pairs of collinear wire segments are linked or remain separated in the ASIC design by including or omitting corresponding local connections from the lower metal layers. Accordingly, interconnect wires of varying lengths are selectively included in an ASIC design, with each long-distance interconnection being formed by linking two or more optimal-length wire segments. Because the segmentation and buffer sizing are pre-engineered to pass signals in minimum time, all long-distance signal transmissions will reach their destination with a minimum delay possible (except for some additional loading from-parts of the first and last segments that extend past the source and destination).
In accordance with yet another aspect of the present invention, source and destination circuits associated with a long-distance signal transmission are placed such that the input terminal of an interconnect buffer is used as a driver for the source of the signal. Selectively positioning the source circuit is made possible because the lower level metal interconnect and transistor layers are customized for each ASIC design.
In accordance with yet another aspect of the present invention, signal branching (i.e., fan-in or fan-out) is accomplished by several methods. First, where a wire branches to two destinations, this branch can be limited to the intersection of the segments, with both downstream branch segments being driven by the same buffer. Alternatively, local interconnection can be provided to link the upstream wire segment to two buffers, each driving one of the downstream branches. In this alternative embodiment, signals may be routed from the end of a long-distance segment, through short-distance routing into the buffer for the following long-distance segment. In yet another embodiment, wires are not permitted to branchxe2x80x94each wire from a source to a destination must have its own wiring path. This is very inefficient in area, particularly for high fan-out signals, but provides good wire delay estimation. In another embodiment, connections may tap off long-distance wires directly, though this introduces delays into the tapped wire segment. Alternatively, connections may be buffered off the long-distance interconnect to preserve the optimal-delay property of long-distance interconnect.